Fall Detection Using LBFD-Net: A Novel CNN-Based Architecture

The aging population has made elderly care a major public health concern, particularly regarding the risk of fall accidents. This paper presents our Lightweight Binary Fall Detection Network (LBFD-Net), designed to address the challenges of fall detection. LBFD-Net uses depthwise separable convolutions, which significantly reduce computational cost while maintaining strong spatial feature extraction. Our architecture has been evaluated using a custom dataset composed of URFD, UP-Fall, Le2i, and MultiCam, consisting of 9,642 (≈52.64%) frames of falls and 8,674 (≈ 47.36%) frames of activities of daily living (ADL) sequences. It achieved an average accuracy of 99.59% and an F1-score of 99.61%, performing better than LeNet and AlexNet architectures adapted for binary fall detection. The source code, pre-trained weights, and the merged dataset are available at https://github.com/nmahammad/LBFD-Net.